← All notes & connection map

Large HPC Systems

Created on: 2025-02-16
Tags: #LargeHPCSystems
Course: HPC


Definition

  • Massively Parallel Processors
  • Distributed memory at largest scale. Often shared memory at lower hierarchies.
  • Fastest SC - Frontier

Comparison of CPU v/s GPU

  • A CPU has several very powerful ALU but seen as a single core.
  • It as a large cache.
  • CPUs can calculate very complicated calculations by themselves without breaking them down into sub-instructions.
  • CPU works on SISD architecture whereas GPU operate on the SIMD architecture.
  • GPUs have multiple lightweight CPUs which can perform a single simple task parallelly on multiple data.

Analogy: It is more effective to deliver Pizza’s through light duty scooters rather than big trucks. Similarly effective to use several lightweight GPU processors for parallel tasks.

  • Performance share of accelerators/GPU is increasing day-by-day.

Frontier System

  • 9408 compute nodes across 74 cabinets.
  • AMD 64-core 3rd generation CPUs
  • AMD Intinct MI250X GPUs
  • Connected through high-speed HPE Slingshot interconnect.
  • 679 PB storage

Node Design

  • 2x nodes per blade.
  • Direct liquid cooled.
  • 1 Shared Memory Machine
  • Each node has
    • 1 CPU
    • 4 AMD GPU
    • 512 RAM
    • 512 GB High Bandwidth Memory
    • 2x 1.92 TB NVMe
    • 4x HPE Slingshot 200 GbE NICs

AMD Instinct MI250X GPU

  • 2 Graphics Compute Dies per GPU.
  • Shown by OS as 2 GPUs/
  • Effectively 8 GPU per node.
  • 26.5 TFLOPs double-precision peak.
  • 110 compute units.
  • 64 GB of HBM
  • 1.6 TB/s Memory Bandwidth

Linked from

  • No other published notes link here yet.